BaCon: Boosting Imbalanced Semi-supervised Learning via Balanced Feature-Level Contrastive Learning
Authors: Qianhan Feng, Lujing Xie, Shijie Fang, Tong Lin
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method demonstrates its effectiveness through comprehensive experiments on the CIFAR10-LT, CIFAR100LT, STL10-LT, and SVHN-LT datasets across various settings. |
| Researcher Affiliation | Collaboration | Qianhan Feng1,2, Lujing Xie3, Shijie Fang2,4*, Tong Lin1,2 1National Key Laboratory of General Artificial Intelligence, China 2School of Intelligence Science and Technology, Peking University 3Yuanpei College, Peking University 4Google, Shanghai, China |
| Pseudocode | No | No structured pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | No | The paper states 'We implement all the algorithms based on USB (Wang, Chen, and Fan 2022) framework' but does not explicitly provide a link or statement about the availability of their own source code for the proposed method. |
| Open Datasets | Yes | We first construct imbalance datasets based on several benchmark datasets including CIFAR10, CIFAR100 (Krizhevsky, Hinton et al. 2009), STL10 (Coates, Ng, and Lee 2011) and SVHN (Netzer et al. 2011). |
| Dataset Splits | No | The paper describes how imbalanced training datasets are constructed based on benchmark datasets and evaluation is done on a class-balanced test set, but it does not explicitly provide specific details for a separate validation dataset split. |
| Hardware Specification | Yes | We implement all the algorithms based on USB (Wang, Chen, and Fan 2022) framework and use a single RTX 3090 GPU to train models. |
| Software Dependencies | No | The paper states 'We implement all the algorithms based on USB (Wang, Chen, and Fan 2022) framework' but does not provide specific version numbers for programming languages or other critical software libraries. |
| Experiment Setup | Yes | SGD is used to optimize parameters. Each mini-batch includes 64 labeled samples and 64 uratio unlabeled samples, and uratio varies for different base SSL algorithms. The learning rate is initially set as η0 = 0.03 with a cosine learning rate decay schedule as η = η0 cos ( 7πt 16T ). The total number od training for each algorithm is 300,000, and the first 100,000 is warmup stage by default. |